MLU100-C Series Intelligent Processing Card User Manual Cambricon, MLU100-C Series Intelligent Processing Card User Manual, 2018-09 Revision History Document Name MLU100-C Series Intelligent Processing Card User Manual Revision Created Date Revision No. V1.0.1 2018-09-01 0 1 2 3 4 Draft Date 2018.09.01 2018.09.06 2018.10.22 2018.12.1 Revision No. Notes:
V1.0.0 V1.0.1 V1.0.2 V1.0.3 First Release Power consumption, heat dissipation and power supplies updated Peak Processing Capacity and PCIe BAR updated Coding/decoding related descriptions updated. Technical supports contact: service@cambricon.com www.cambricon.com Cambricon, MLU100-C Series Intelligent Processing Card User Manual, 2018-09 Content 1. Product Brief ..........................................................................................................................................4 2. Product specifications...........................................................................................................................5 Performance specifications .....................................................................................................5 2.1 Software specifications ...........................................................................................................6 2.2 2.3 Working Environment Specifications ...................................................................................7 Outline Dimension Specifications .........................................................................................7 2.4 Heat Dissipation Specifications .............................................................................................7 2.5 Power Supplies Specifications ...............................................................................................8 2.6 3. Software Development Environment .............................................................................................. 11 4. FCC notes ............................................................................................................................................ 11 Cambricon, MLU100-C Series Intelligent Processing Card User Manual, 2018-09 1. Product Brief Figure. 1 MLU100-C Series Intelligent Processing Card appearance Cambricon MLU100-C Series Intelligent Processing Card The MLU100-C Series is the third generation of Intelligent Processing Cards provided by Cambricon Technologies, on which populated one piece of Cambricon MLU100 chip. It can provide extremely high calculating and video coding/decoding capacities to the Cloud Inference. Compared with classical general-purpose CPU, MLU100-C can achieve a much bigger Performance/Power-
consumption ratio when handling the AI (Artificial Intelligence) related works, which makes it a real AI-ready processor. MLU100-C Series Intelligent Processing Card has a theoretical equivalent Peak Processing Capacity of about 128TOPS(INT8), and it can be integrated into a server or workstation by a PCIe Gen3 x16 single slot. The Cards heat dissipation solution is passive so the server (workstation) system should be in charge of providing cooling airflow. The typical power consumption of the card is about 60-90W. MLU100-C Series Intelligent Processing Card can support up to 16GB DDR4 with ECC capacity. Cambricon MLU100-C Series Intelligent Processing Cards key features:
1. Versatility As a universal intelligent processor, it can support many kinds of Deep Learning technologies and multimodal intelligent processing such as vision, voice and nature language, and it can be used in a wide range. 2. Sparsification By bringing sparsity technology into AI chip, it can achieve a theoretical peak performance of 128TOPS (INT8). 3. High efficiency Compared with GPU, it has featured special Instruction Set and processor architecture dedicated to deep learning and AI, so it has much higher Performance/Power-
consumption ratio. 4. Perfect software development environment Cambricon Technologies provides a thorough and Cambricon, MLU100-C Series Intelligent Processing Card User Manual, 2018-09 mature development environment Cambricon NeuWare for MLU100-C series processors, and it has included a full set of tools such as Application Development, Function Debugging and Performance Optimizations. 5. CE, FCC, UL certificates. 2. Product specifications 2.1 Performance specifications Table 1 describes the performance specifications of Cambricon MLU100-C Series Intelligent Processing Card. Card Type Processor Architecture Core Clock Peak FP16 Peak INT8 Video Coding/Decoding Support Total Memory Size Memory Bus Width Memory Bandwidth System Interfaces PCIe ID Outline Typical Power TDP ECC Supporting MLU100-C3/C4 Cambricon MLUv01 1 GHz 16 TFLOPS (Without Sparsity) 64 TFLOPS (With Sparsity) 32 TOPS (Without Sparsity) 128 TOPS (With Sparsity) Yes 8GB/16GB 256-bit 102.4 GB/s PCI Express Gen3, x16, Supporting Lane Reversal PCIe Vendor ID PCIe Device ID PCIe Sub-Vendor ID 0xCABC PCIe Sub-System ID 0x00C3/0x00C4 267mm(L)*111.15mm(W), Single-Slot 60-90W 110W Yes 0xCABC 0x0100 Table 1 PCIe Card hardware Specifications Cambricon, MLU100-C Series Intelligent Processing Card User Manual, 2018-09 2.2 Software specifications Table 2 Describes the software specifications of MLU100-C Series Card PF 164bit:
BAR0: 256MB prefetchable BAR2: 16MB prefetchable BAR4: 64MB prefetchable VF464bit BAR0: 16MB prefetchable BAR2: 16MB prefetchable BAR4: 16MB prefetchable Yes (Enabled by default) 0x8E (write) 0x8F (read) PCIe Base addressMLU100 ECC Protect SMBus (8bit Address) Table 2 PCIE Card Software Specifications SMBUS Register is 32-bit wide, and Table 3 describes how to read a register:
DIRECTION M->S BITS CONTENT 1 S M->S 8 SLAVE ADDRESS Write S->M 1 M->S 8 S->M 1 ACK REGISTER ADDRESS ACK M->S M->S 1 Sr 8 SLAVE ADDRESS Read S->M S->M M->S S->M M->S S->M M->S S->M M->
S 1 N 1 8 1 8 1 8 1 8 ACK DATA[
7:0]
ACK DATA[
15:8]
ACK DATA[
23:16]
ACK DATA[
31:24]
M->
S 1 P Table 3 SMBUS register reading Table 4 describes the definition and address of SMBUS registers:
Registers Total Card Power Card Temperature Chip Temperature PCIE Vendor ID and Device ID PCIE Sub-Vendor ID and Sub-System ID Device Name Address Access 0x01 0x02 0x03 0xA0 RO RO RO RO 0xA1 0xF0 RO RO Description Card Power consumption, Float Data, Unit W Card Temperature, Float Data, Unit Chip Temperature, Float Data, Unit
[15:0] Vendor ID
[31:16] Device ID
[15:0] Sub-Vendor ID
[31:16] Sub-System ID Cambricon, MLU100-C Series Intelligent Processing Card User Manual, 2018-09 Vendor Name Hardware Revision Firmware Revision Manufacturing Time Device ID 0xF1 0xF2 0xF3 0xF4 0xF5 RO RO RO RO RO Table 4 SMBUS Registers Description 2.3 Working Environment Specifications Operating Temperature Storage Temperature Operating Humidity Storage Humidity 045C
-4075 5%95% Relative Humidity 5%95% Relative Humidity Table 5 PCIE Card Working Environment 2.4 Outline Dimension Specifications Figure 2 MLU100-C outline dimension 2.5 Heat Dissipation Specifications MLU100-C Series Intelligent Processing Card is using a passive heat dissipation method. The system designers should pay attentions to below for the Airflow Volume at the entrance and Temperature of Ambient relationship, so as to design a cooling system for the card and guarantee the card is operating within the green recommendation area. The core clock frequency of the chip will be decreased to half when the temperature of chip is over 87, and shut down when over 90. Cambricon, MLU100-C Series Intelligent Processing Card User Manual, 2018-09 Figure 3 Temperature of Ambient vs Airflow Volume Below is the supported direction of airflow: It supports either of directions. Figure 4 Supported direction of airflow 2.6 Power Supplies Specifications Refer to below, MLU100-C Series Intelligent Processing Card provides an 8-pin CPU Power Supply Socket:
Cambricon, MLU100-C Series Intelligent Processing Card User Manual, 2018-09 Figure 5 8-Pin CPU Power socket Users can use the 8-Pin CPU Power Socket within a Server directly, or use the 8-Pin PCIe Power Socket, along with a One-to-Two cable. Refer to below for a One-To-Two cable:
Figure 6 One-To-Two Cable 12V 8-Pin CPU Socket 5 8-Pin PCIe Socket 1 3 8-Pin PCIe Socket 2 Color Yellow Cambricon, MLU100-C Series Intelligent Processing Card User Manual, 2018-09 12V 12V 12V GND GND GND GND 6 7 8 1 2 3 4 1,2 7,8 5,6 3 1,2 7,8 5,6 Yellow Yellow Yellow Black Black Black Black Table 6 Cable Wires relationship Figure 7 Adapter Cable Connection Relationship Note: The first pin (Pin No.1) of every sockets should have an obvious distinguishable mark at the side. Attentions should be paid to it. Power Supply Min (V) Normal(V) Max(V) PCIe edge connecter12V PCIe 8-pin connecter12V PCIe edge connecter3V3 10.8 10.8 3.0 12 12 3.3 13.2 13.2 3.6 Current Avg (A) Current Peak (A) 2 7.5 0.2A 3 11.25 0.3A Table 7 Power Supplies - Voltages and Currents Specifications Cambricon, MLU100-C Series Intelligent Processing Card User Manual, 2018-09 3. Software Development Environment Cambricon Technologies provides a software development environment: Cambricon NeuWare NeuWare can fully support all kinds of mainstream programming framework, such as TensorFlow, Caffe, Caffe2, MXNet and ONNX. With above mentioned Programming Frameworks, users can easily and conveniently develop and deploy their Deep Learning Applications on Cambricon MLU100-C Series Intelligent Processing Cards. At the same time, NeuWare provides complete runtime system and driver software to speed up the system integration procedure. NeuWare also provides a full set of software tools such as Application Development, Function Debugging and Performance Optimizations. The Application Development Tools include Machine Learning Library, Runtime Library, Compiler, Model retraining tools and Domain-Specific (for example Video Analysis) SDK; The Function Debugging Tools can fulfill all the requirements from different levels of Programming Framework and Function Library; The Performance Optimization Tools include tools for performances analysis and system monitoring. Figure 8 Cambricon NeuWare For more information, please refer to: www.cambricon.com Tel: +86-10-83030003 Email: business@cambricon.com Address: 11 Floors, Block D, Zhizhen Building, 7 Zhichun Road, Haidian District, Beijing 4. FCC notes This device complies with Part 15 of the FCC Rules. Operation is subject to the following two conditions: (1) This device may not cause harmful interference, and (2) this device must accept any interference received, including interference that may cause undesired operation. NOTE: This equipment has been tested and found to comply with the limits for a Class B digital device, pursuant to Part 15 of the FCC Rules. These limits are designed to provide reasonable protection against harmful interference in a residential installation. This equipment generates, uses and can radiate radio frequency energy and, if not installed in accordance with the instructions, may cause harmful Cambricon, MLU100-C Series Intelligent Processing Card User Manual, 2018-09 interference to radio communications. However, there is no guarantee that interference will not occur in a particular installation. If this equipment does cause interference to radio or television reception, which can be determined by turning the equipment off and on, the user is encouraged to try to correct the interference by one or more of the following measures:
Reorient or relocate the receiving antenna Increase the separation between the equipment and receiver Connect the equipment into an outlet on a circuit different from that to which the receiver is connected Consult the dealer or an experienced radio/TV technician for help Information User:
Changes or modifications not expressly approved by the party responsible for compliance could void the user's authority to operate the equipment. Cambricon, MLU100-C Series Intelligent Processing Card User Manual, 2018-09